The oracle for hidden side effects was in the fixtures already

D129 said nothing here could observe whether an instruction writes the condition code - it is invisible in the encoding, the operand layout, and any test that checks destinations. That was wrong, and the counter-example had been sitting in control.txt since it was generated:

s_cmp_lt_i32   <- sets it
s_mov_b64      <- between
s_cbranch_scc1 <- branches on it

A compiler putting an instruction there is asserting it does not write the condition code, or the shader it emitted would be wrong. the_corpus_agrees_about_hidden_side_effects mines every fixture for those windows and checks them against the translator. One window, one instruction confirmed, and it grows with the corpus.

246 tests green across the shader-side crates.

Surprises

• "No oracle exists" was a claim about attention, not about the world. The evidence was in a file this project generated, in a test directory it reads on every run. What was missing was the question.
• A list is not automatically the D122 mistake. That one duplicated derivable data. This one encodes a fact with a single source and no derivation - the fault there was the duplication, not the listing, and the fix here is a check rather than a deletion.